Moses-support Digest, Vol 121, Issue 25

Send Moses-support mailing list submissions to
moses-support@mit.edu

To subscribe or unsubscribe via the World Wide Web, visit
http://mailman.mit.edu/mailman/listinfo/moses-support
or, via email, send a message with subject or body 'help' to
moses-support-request@mit.edu

You can reach the person managing the list at
moses-support-owner@mit.edu

When replying, please edit your Subject line so it is more specific
than "Re: Contents of Moses-support digest..."


Today's Topics:

1. Hyerarchial output search graph (Guillem Torres Badia)
2. Character ngrams using KenLM (Nat Gillin)
3. Re: Ensemble of Neural Machine Translation systems (Nat Gillin)
4. Re: Character ngrams using KenLM (Kenneth Heafield)
5. Re: Ensemble of Neural Machine Translation systems (Rico Sennrich)


----------------------------------------------------------------------

Message: 1
Date: Tue, 08 Nov 2016 23:21:31 +0100
From: Guillem Torres Badia <guitorba@inf.upv.es>
Subject: [Moses-support] Hyerarchial output search graph
To: moses-support@mit.edu
Message-ID:
<20161108232131.Horde.ZFa_acmUv4TWYP64SKX9z0i@webmail.upv.es>
Content-Type: text/plain; charset="utf-8"

Hello,

I am training an output-search-graph/hypergraph with MOSES chart. I
would like to know which are the training parameters I should consider
and their values that usually give best results in BLEU.

I attach you the script I'm using currently, just in case you need to
check some details.

Guillem
-------------- next part --------------
A non-text attachment was scrubbed...
Name: launch_es-en.sh
Type: application/x-sh
Size: 2588 bytes
Desc: not available
Url : http://mailman.mit.edu/mailman/private/moses-support/attachments/20161108/a993aacd/attachment-0001.sh

------------------------------

Message: 2
Date: Wed, 9 Nov 2016 14:04:07 +0800
From: Nat Gillin <nat.gillin@gmail.com>
Subject: [Moses-support] Character ngrams using KenLM
To: moses-support@mit.edu
Message-ID:
<CAD2EOZjhaf1XwuSp-FPZqmUudDAJw+=7rPYkPT1h4a5x51cCYw@mail.gmail.com>
Content-Type: text/plain; charset="utf-8"

Dear Moses community,

Other than manually replacing space with an unused character and adding
spaces to each character before training a language model with KenLM. Is it
possible for KenLM to generate character ngrams and output in arpa format
without altering the input file?

Regards,
Nat
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://mailman.mit.edu/mailman/private/moses-support/attachments/20161109/83e81335/attachment-0001.html

------------------------------

Message: 3
Date: Wed, 9 Nov 2016 15:01:36 +0800
From: Nat Gillin <nat.gillin@gmail.com>
Subject: Re: [Moses-support] Ensemble of Neural Machine Translation
systems
To: moses-support@mit.edu
Message-ID:
<CAD2EOZj2tdZut+ZrqqBrmW+=5rh-uYi5RmCkxEr4=h4705igbg@mail.gmail.com>
Content-Type: text/plain; charset="utf-8"

Dear Rico and Moses community,

Thanks for the response. Is there a reason why should there be an average
at every time step? Is the assumption that all networks contributing to the
ensembles are equal?

How is that different from training wider neural nets?

Sorry for the multiple questions.

Regards,
Nat
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://mailman.mit.edu/mailman/private/moses-support/attachments/20161109/bec5185c/attachment-0001.html

------------------------------

Message: 4
Date: Wed, 09 Nov 2016 08:46:27 +0000
From: Kenneth Heafield <moses@kheafield.com>
Subject: Re: [Moses-support] Character ngrams using KenLM
To: Nat Gillin <nat.gillin@gmail.com>, moses-support@mit.edu
Message-ID: <2A50A96D-1591-4B6B-BF44-BCA41F394E2D@kheafield.com>
Content-Type: text/plain; charset="utf-8"

No. Tokenizer and LM are separate tools. You can of course replace space with a token like <spc> or something.

On November 9, 2016 6:04:07 AM GMT+00:00, Nat Gillin <nat.gillin@gmail.com> wrote:
>Dear Moses community,
>
>Other than manually replacing space with an unused character and adding
>spaces to each character before training a language model with KenLM.
>Is it
>possible for KenLM to generate character ngrams and output in arpa
>format
>without altering the input file?
>
>Regards,
>Nat
>
>
>------------------------------------------------------------------------
>
>_______________________________________________
>Moses-support mailing list
>Moses-support@mit.edu
>http://mailman.mit.edu/mailman/listinfo/moses-support
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://mailman.mit.edu/mailman/private/moses-support/attachments/20161109/f63fb288/attachment-0001.html

------------------------------

Message: 5
Date: Wed, 9 Nov 2016 13:27:19 +0000
From: Rico Sennrich <rico.sennrich@gmx.ch>
Subject: Re: [Moses-support] Ensemble of Neural Machine Translation
systems
To: moses-support@mit.edu
Message-ID: <88d59b3b-c230-2489-7a45-faf21c41bbb5@gmx.ch>
Content-Type: text/plain; charset="windows-1252"

Hi Nat,

The reason for averaging at every time step (rather than doing k-best
list reranking on the sentence level) is the same reason why we
integrate new feature functions in Moses instead of just reranking the
k-best output: you make more search errors if you do your initial search
with a weak model, and then re-rank the k-best list with a stronger
model, and it is sensible to do the full search with the stronger model.

you can also do a weighted average if you want.

best wishes,
Rico


On 09/11/16 07:01, Nat Gillin wrote:
> Dear Rico and Moses community,
>
> Thanks for the response. Is there a reason why should there be an
> average at every time step? Is the assumption that all networks
> contributing to the ensembles are equal?
>
> How is that different from training wider neural nets?
>
> Sorry for the multiple questions.
>
> Regards,
> Nat
>
>
> _______________________________________________
> Moses-support mailing list
> Moses-support@mit.edu
> http://mailman.mit.edu/mailman/listinfo/moses-support

-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://mailman.mit.edu/mailman/private/moses-support/attachments/20161109/11264563/attachment.html

------------------------------

_______________________________________________
Moses-support mailing list
Moses-support@mit.edu
http://mailman.mit.edu/mailman/listinfo/moses-support


End of Moses-support Digest, Vol 121, Issue 25
**********************************************

Tweet